13 priors, 6 model templates, and 9 open-source PFN projects to fork, learn from, or build on.
priors, models, and external projects
The reference PFN prior. Random linear functions with Gaussian noise — the simplest demonstrable PFN training task.
Classification analogue of linear regression. Each task is a fresh draw of the two-moons geometry — the PFN learns a generic 2D classifier.
Real-world-shaped time series. Random stationary AR(2) coefficients per task; the PFN learns to forecast any well-behaved autoregressive series.
Random sparse linear structural causal models. The PFN outputs an adjacency matrix — pure structure discovery, no fitted edges.
Random polynomial functions up to degree D. Generalises linear regression with curvature.
Functions sampled from a Gaussian Process with an RBF kernel. PFN learns to do GP regression at inference without solving the kernel system.
Random Gaussian mixtures in D dimensions. Trains a PFN that does Bayesian-optimal classification on any well-separated mixture.
Logistic regression with pairwise feature interactions baked in. Trains a PFN that picks up cross-feature signal automatically.
The simplest temporal PFN. Random amplitude, frequency, and phase per task.
Classic decomposition: slow trend plus periodic component plus noise. Demo-worthy for retail / energy / web traffic.
Textbook conjugate prior. Direct evidence the PFN has learned Bayesian inference.
Two-level normal model. Groups share information through a population mean — the canonical multi-level setup.
Specialised SCM prior with chain-shaped DAGs. Faster to learn than full ER-DAGs and matches a common scientific use case.
Smallest model that consistently solves linear-regression-shape priors. Good first pass for any 1D scalar-output task.
Same backbone as the regression baseline, scalar head emits one logit per point. Paired with the two-moons / GMM priors.
Deeper backbone for sine / AR / seasonal priors. Wider d_model and four attention layers give enough capacity for non-trivial dynamics.
Outputs a d×d adjacency matrix from N×d observations. The default architecture for linear-SCM / chain-SCM discovery priors.
Wider, deeper tabular backbone for harder priors (hierarchical models, GMM with many classes, interaction-heavy logistic).
For causal effect estimation tasks where the output is a real-valued estimate per query. Pairs with potential-outcome priors.
AutoML Freiburg's maintained PFN library — the canonical implementation for training transformer-based PFNs that approximate Bayesian prediction. Foundation for TabPFN, PFNs4BO, LC-PFN and most downstream PFN work.
In-context-learning transformer that predicts on small tabular datasets in seconds, no per-dataset training. v2 was published in Nature (2025) and matches or beats tuned tree ensembles.
Frames forecasting as tabular regression and runs TabPFN v2 with lightweight feature engineering for zero-shot point + probabilistic forecasts. Handles exogenous features (weather, holidays) without preprocessing.
Foundational hypernetwork trained on synthetic tabular tasks: prompted with a training set, it emits the weights of a small child neural network without gradient descent. Faster inference than TabPFN.
ICML 2023 implementation using PFNs as surrogates for Bayesian optimisation, replacing Gaussian processes with a pre-trained transformer that predicts posteriors in one forward pass.
ICML 2024. Uses a PFN as a freeze-thaw surrogate, predicting learning-curve continuations to decide which configurations to keep training. Anytime-efficient hyperparameter search.
NeurIPS 2023. Predicts the posterior over a learning curve's continuation given a few initial points, using a PFN trained on a parametric curve prior. Drop-in surrogate for early-stopping.
Permissive BSD-licensed tabular foundation model in the PFN family. Strong benchmark results plus a forecast sub-module derived from TabPFN-TS.
ICLR 2025. PFN applied to RNA kinetics: a transformer trained on synthetic kinetic priors directly predicts the CDF of first passage times, replacing expensive Kinfold simulations.
